EZ-SP: Lightweight GPU 3D Segmentation

• EZ-SP is a GPU-based 3D semantic segmentation system that replaces CPU-bound superpoint partitioning with a learnable, parallelized approach.
• It employs a three-stage pipeline—point embedding, GPU superpoint partitioning via graph optimization, and hierarchical classification—to achieve rapid inference.
• Experiments on S3DIS, KITTI-360, and DALES show up to 72× faster inference and significantly fewer parameters compared to traditional point-based models.

Searching arXiv for “EZ-SP” to confirm the relevant paper and possible disambiguations. arXiv search query: EZ-SP EZ-SP, expanded as “easy-superpoints,” is a fast, lightweight, fully GPU-based 3D semantic segmentation pipeline that replaces the usual CPU-bound superpoint partitioning stage with a learnable partitioner and then performs classification at the superpoint level rather than on all individual points or voxels (Geist et al., 29 Nov 2025). In this formulation, a raw point cloud is segmented into superpoints on the fly, and a compact superpoint-level classifier predicts semantic labels for those regions. The method is positioned as an answer to a specific systems bottleneck in prior superpoint-based pipelines: partitioning was often CPU-bound, dependent on handcrafted geometric or radiometric features, slow to tune, and difficult to scale to very large scenes (Geist et al., 29 Nov 2025).

1. Terminology and scope

In the arXiv record, EZ-SP most directly denotes the 3D semantic segmentation system introduced in “EZ-SP: Fast and Lightweight Superpoint-Based 3D Segmentation” (Geist et al., 29 Nov 2025). The acronym is not globally unique. The string also appears in unrelated contexts: EZ Lyn is referred to in the text as EZ-SP in the context of an accreting white dwarf pulsator (Pavlenko et al., 2014), and UC2-ESP is described as the concrete implementation if EZ-SP refers to an ESP32-based microscope-control idea or tool (Diederich et al., 26 Oct 2025). This suggests that, in technical usage, EZ-SP should be disambiguated by domain.

Within 3D vision, however, EZ-SP has a precise meaning. It is a superpoint-based segmentation system that makes the partition stage learned, GPU-native, compact, and fast, while preserving the computational advantage of operating on regions instead of points or voxels (Geist et al., 29 Nov 2025). The central claim is not merely that superpoints are useful, but that they can be made practical at scale once the CPU partition bottleneck is removed.

2. Pipeline architecture

EZ-SP consists of three main stages: point embedding or semantic transition detection, GPU superpoint partitioning, and superpoint classification followed by dense label broadcasting (Geist et al., 29 Nov 2025). The pipeline first learns a low-dimensional embedding for each point, then clusters those embeddings into superpoints, and finally applies a lightweight network that predicts a label per superpoint (Geist et al., 29 Nov 2025).

The embedding network is a very small sparse CNN implemented with TorchSparse, using 3 layers of width $[32,32,32]$ and embedding dimension $M=32$ (Geist et al., 29 Nov 2025). Its output is the pointwise embedding matrix

$$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$

where $\mathcal{C}$ is the point cloud and $\mathbf{f}_p$ is the embedding of point $p$ (Geist et al., 29 Nov 2025). The intended role of this encoder is not direct pointwise semantic classification. Instead, it is trained so that semantic boundaries become separable in the learned embedding space.

This design is motivated by the observation that semantic boundaries usually coincide with sharp geometric changes and radiometric changes such as color or intensity shifts (Geist et al., 29 Nov 2025). A plausible implication is that EZ-SP treats boundary detection as the upstream primitive from which coherent superpoints can be derived, rather than attempting to solve the entire semantic task at point granularity.

3. Boundary-aware embedding and differentiable training objective

To train the embeddings, EZ-SP uses a graph-based binary edge-classification formulation on an undirected $k$-NN graph $(\mathcal{C},\mathcal{E})$ (Geist et al., 29 Nov 2025). Edges are divided into intra-edges,

$\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$

and inter-edges,

$\mathcal{E}_{\text{inter}}=\{(p,q)\in\mathcal{E}\mid \class(p)\neq\class(q)\},$

so that the learned embedding is explicitly optimized to distinguish within-class from cross-class neighborhood relations (Geist et al., 29 Nov 2025).

For points $M=32$0 and $M=32$1, EZ-SP defines the affinity

$M=32$2

with temperature $M=32$3, and in experiments $M=32$4 (Geist et al., 29 Nov 2025). The training loss enforces $M=32$5 on intra-edges and $M=32$6 on inter-edges through

$M=32$7

(Geist et al., 29 Nov 2025). The paper describes this as a differentiable surrogate for boundary detection.

Because intra-class edges can dominate, the method applies adaptive sampling: intra-edges are randomly dropped, with at most a fraction $M=32$8, taking reported values $M=32$9 for S3DIS and $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$0 for KITTI-360 and DALES (Geist et al., 29 Nov 2025). The transition module is trained with Adam, learning rate $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$1 on S3DIS, learning rate $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$2 on KITTI-360 and DALES, weight decay $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$3, and focal loss with $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$4 (Geist et al., 29 Nov 2025). The module has under 60k parameters and trains in under 20 minutes (Geist et al., 29 Nov 2025).

One common misconception is that EZ-SP is simply a classical clustering front-end attached to a segmentation head. That characterization is incomplete. The partition is driven by learned point embeddings optimized specifically for semantic transitions, and the partition step itself is formulated as a GPU-parallel optimization rather than as a fixed handcrafted preprocessor (Geist et al., 29 Nov 2025).

4. Superpoint partitioning as GPU graph optimization

The learned embeddings are converted into superpoints through a contour-regularized graph optimization. EZ-SP seeks a piecewise-constant approximation $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$5 of the embeddings $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$6 by minimizing

$$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$7

(Geist et al., 29 Nov 2025). The first term preserves fidelity to the embeddings, while the second encourages piecewise-constant regions. In the reported experiments, $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$8 and $$\mathbf{F} = \phi^{\text{point}}(\mathcal{C}) \in \mathbb{R}^{|\mathcal{C}| \times M},$$9 (Geist et al., 29 Nov 2025).

Rather than solving this objective with a CPU-bound solver, EZ-SP converts it into a combinatorial merging problem over connected components (Geist et al., 29 Nov 2025). For a current superpoint $\mathcal{C}$0, the component embedding is

$\mathcal{C}$1

and the partition induces superpoint adjacency

$\mathcal{C}$2

(Geist et al., 29 Nov 2025). If adjacent superpoints $\mathcal{C}$3 and $\mathcal{C}$4 are merged, the energy decrease is

$\mathcal{C}$5

(Geist et al., 29 Nov 2025). A merge is beneficial when $\mathcal{C}$6, unless the minimum size constraint forces merging (Geist et al., 29 Nov 2025).

The GPU partition algorithm is a parallel greedy bottom-up merging method implemented with CUDA-accelerated PyTorch operations and scatter-based primitives (Geist et al., 29 Nov 2025). It initializes with each point as its own superpoint,

$\mathcal{C}$7

then builds directed merge edges $\mathcal{C}$8 whenever $\mathcal{C}$9 or $\mathbf{f}_p$0, restricts each source component to one outgoing edge with highest $\mathbf{f}_p$1, computes weakly connected components of the directed merge graph, merges all nodes in each component at once, and repeats until no valid merges remain (Geist et al., 29 Nov 2025). The appendix provides a recursive max-propagation algorithm for computing weakly connected components efficiently on GPU (Geist et al., 29 Nov 2025).

The critical engineering idea is that conflict removal permits parallelism: each source superpoint selects only one outgoing merge edge, which removes merge conflicts and allows the merge graph to be processed in parallel (Geist et al., 29 Nov 2025). This suggests that EZ-SP’s speedup is not only a matter of moving computation from CPU to GPU, but also of reformulating the merge schedule to fit GPU execution.

5. Hierarchical representation and superpoint classification

EZ-SP can be applied recursively to form a hierarchy

$\mathbf{f}_p$2

where each $\mathbf{f}_p$3 partitions the components of $\mathbf{f}_p$4 (Geist et al., 29 Nov 2025). The reported minimum superpoint sizes are $\mathbf{f}_p$5 for S3DIS and KITTI-360, and $\mathbf{f}_p$6 for DALES (Geist et al., 29 Nov 2025). The hierarchy is not incidental: the downstream classifier benefits from multi-scale structure, and removing the third hierarchical level reduces S3DIS Fold 5 mIoU from $\mathbf{f}_p$7 to $\mathbf{f}_p$8 while leaving throughput about the same (Geist et al., 29 Nov 2025).

For semantic segmentation, EZ-SP uses a modified SuperPoint Transformer. The reported instantiations are SPT-64 on S3DIS and DALES and SPT-128 on KITTI-360 (Geist et al., 29 Nov 2025). The modifications include removing all CPU-bound preprocessing, using the new GPU-generated hierarchical partitions, preserving most of the SPT design, adding a third hierarchical stage, reintroducing the feed-forward layer in the transformer block for DALES, and concatenating color, position, and elevation with CNN features to build point embeddings (Geist et al., 29 Nov 2025). The model predicts a label per superpoint and then broadcasts that label back to the constituent points (Geist et al., 29 Nov 2025).

The segmentation loss is focal loss, with $\mathbf{f}_p$9 for S3DIS and $p$0 for KITTI-360 and DALES (Geist et al., 29 Nov 2025). Reported classifier sizes are 330k parameters on S3DIS, 870k on KITTI-360, and 425k on DALES (Geist et al., 29 Nov 2025). The paper also notes that, for the segmentation head, the first MLP layer is widened from $p$1 to $p$2 to match the richer learned embeddings (Geist et al., 29 Nov 2025).

A second misconception is that superpoint methods are lightweight only at classification time while retaining a heavy preprocessing dependency. EZ-SP explicitly targets that historical weakness by removing CPU-bound preprocessing and integrating partition generation into the GPU pipeline (Geist et al., 29 Nov 2025).

6. Empirical performance, scaling, and efficiency

EZ-SP is evaluated on three benchmarks: S3DIS, with 273M points and 13 classes in the merged floor setting; KITTI-360, with 919M points, 15 classes, and 61 validation scans; and DALES, with 492M points and 8 classes over 12 evaluation scans (Geist et al., 29 Nov 2025). Subsampling uses a 3 cm grid for S3DIS and a 10 cm grid for KITTI-360 and DALES (Geist et al., 29 Nov 2025). Optimization uses Adam and a cosine learning-rate schedule with 50 warm-up epochs for segmentation (Geist et al., 29 Nov 2025).

The reported semantic segmentation scores are S3DIS 6-fold mIoU $p$3, S3DIS Area 5 mIoU $p$4, KITTI-360 validation mIoU $p$5, and DALES test mIoU $p$6 (Geist et al., 29 Nov 2025). In the comparison table, these results are close to SPT at $p$7, and are competitive with point-based models such as PointTransformer-v3 at $p$8 on S3DIS 6-fold, PointNeXt-XL at $p$9, and Stratified Transformer at $k$0 (Geist et al., 29 Nov 2025). The paper’s interpretation is that EZ-SP achieves a strong accuracy–efficiency trade-off: near-SOTA accuracy with far lower parameter count, far better throughput, and no CPU partition bottleneck (Geist et al., 29 Nov 2025).

The efficiency results are central. The abstract reports superpoints that are generated 13$k$1 faster than prior methods, a full pipeline with 72$k$2 faster inference and 120$k$3 fewer parameters than point-based SOTA models, and a module compact enough to fit the full pipeline in $k$4 MB of VRAM while supporting real-time inference (Geist et al., 29 Nov 2025). The conclusion reports about $k$5 million points/s, and the paper states that this matches the acquisition rate of automotive LiDAR exceeding $k$6M pts/s (Geist et al., 29 Nov 2025). In the main timing table, EZ-SP partitioning is $k$7 GPU-s versus $k$8 GPU-s for SPT partitioning, while the full S3DIS total is about $k$9 GPU-s for EZ-SP versus $(\mathcal{C},\mathcal{E})$0 GPU-s for SPT (Geist et al., 29 Nov 2025).

The following summary table collects the principal reported quantitative characteristics.

Aspect	Reported value	Context
Partition speedup	13$(\mathcal{C},\mathcal{E})$1 faster	Versus prior superpoint methods (Geist et al., 29 Nov 2025)
Inference speedup	72$(\mathcal{C},\mathcal{E})$2 faster	Versus point-based SOTA models (Geist et al., 29 Nov 2025)
Parameter reduction	120$(\mathcal{C},\mathcal{E})$3 fewer parameters	Versus point-based SOTA models (Geist et al., 29 Nov 2025)
Throughput	about $(\mathcal{C},\mathcal{E})$4 million points/s	Conclusion (Geist et al., 29 Nov 2025)
S3DIS 6-fold mIoU	$(\mathcal{C},\mathcal{E})$5	Semantic segmentation (Geist et al., 29 Nov 2025)
KITTI-360 val mIoU	$(\mathcal{C},\mathcal{E})$6	Semantic segmentation (Geist et al., 29 Nov 2025)
DALES test mIoU	$(\mathcal{C},\mathcal{E})$7	Semantic segmentation (Geist et al., 29 Nov 2025)

The paper also reports scenario memory estimates of about $(\mathcal{C},\mathcal{E})$8 GB for an autonomous driving scan, about $(\mathcal{C},\mathcal{E})$9 GB for a building-scale scene, and about $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$0 GB for an aerial survey (Geist et al., 29 Nov 2025). Examples include a single LiDAR sweep on an embedded Jetson, a full S3DIS floor on a consumer GPU, and large aerial scenes on an NVIDIA A40 (Geist et al., 29 Nov 2025). These results indicate that the low parameter count should not be conflated with trivial deployment conditions: large scenes still demand substantial absolute memory, but the pipeline is designed to process them in one pass (Geist et al., 29 Nov 2025).

7. Relation to prior methods, ablations, and limitations

EZ-SP is situated against both point- or voxel-based networks and earlier superpoint systems. Point and voxel methods such as PointNet++, KPConv, MinkowskiNet, PointNeXt, Stratified Transformer, and PointTransformer-v3 can be highly accurate but are described as expensive, with large parameter counts, high inference latency, heavy test-time augmentation, and limited scalability for huge scans (Geist et al., 29 Nov 2025). The paper specifically notes around $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$1 test-time augmentations for Stratified Transformer and around $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$2 ensembled predictions for PointTransformer-v3 in the reported implementation (Geist et al., 29 Nov 2025). EZ-SP’s counter-position is to avoid that cost by operating on superpoints with a lightweight pipeline.

Against earlier superpoint methods such as SPG, SSP, SPNet, and SPT, the main critique is that classification may be efficient but partitioning dominates runtime because it remains CPU-heavy (Geist et al., 29 Nov 2025). Against VCCS, the critique is rigidity, sensitivity to initialization, and lower adaptability to local scene complexity (Geist et al., 29 Nov 2025). Against PCP, the criticism is that it is CPU-bound, slower, and depends on handcrafted features (Geist et al., 29 Nov 2025). The paper frames EZ-SP as preserving the superpoint advantage while replacing the partition solver with a learned GPU approximation.

The ablations reinforce this interpretation. Replacing the learned CNN with handcrafted geometric features changes S3DIS Fold 5 mIoU only from $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$3 to $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$4 while leaving throughput unchanged at $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$5 M pt/s (Geist et al., 29 Nov 2025). This suggests that the learnable backbone removes the need for feature engineering without sacrificing quality. Disabling GPU or transfer optimizations drops throughput from $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$6 to $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$7 M pt/s (Geist et al., 29 Nov 2025). Replacing the new GPU partitioner with CPU PCP reduces throughput from $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$8 to $\mathcal{E}_{\text{intra}}=\{(p,q)\in\mathcal{E}\mid \class(p)=\class(q)\},$9 M pt/s and lowers accuracy from $\mathcal{E}_{\text{inter}}=\{(p,q)\in\mathcal{E}\mid \class(p)\neq\class(q)\},$0 to $\mathcal{E}_{\text{inter}}=\{(p,q)\in\mathcal{E}\mid \class(p)\neq\class(q)\},$1 (Geist et al., 29 Nov 2025). These ablations make clear that both the learned partition formulation and the implementation strategy are material to the final result.

The paper’s caveats are mostly implicit but concrete. The partitioner still relies on a graph over $\mathcal{E}_{\text{inter}}=\{(p,q)\in\mathcal{E}\mid \class(p)\neq\class(q)\},$2-NN neighborhoods, with $\mathcal{E}_{\text{inter}}=\{(p,q)\in\mathcal{E}\mid \class(p)\neq\class(q)\},$3 nearest neighbors for the partition graph (Geist et al., 29 Nov 2025). Partition quality still depends on how well the transition embeddings capture semantic boundaries (Geist et al., 29 Nov 2025). The method is optimized for semantic segmentation, so extension to other tasks would require adapting the downstream head (Geist et al., 29 Nov 2025). Throughput is reported on standard benchmarks and hardware, and exact speed may vary by implementation and GPU generation (Geist et al., 29 Nov 2025).

Taken together, EZ-SP defines a specific point in the design space of 3D scene understanding: it retains the region-level efficiency of superpoint reasoning, replaces handcrafted and CPU-bound partitioning with a learned GPU-native procedure, and couples that partition with a compact hierarchical classifier (Geist et al., 29 Nov 2025). A plausible broader implication is that EZ-SP repositions superpoint segmentation from a preprocessing-heavy niche to a viable systems-level alternative for large-scale, real-time 3D perception.